Feature engineering is a critical step in addressing the “curse of dimensionality” and high noise levels inherent in laser-induced breakdown spectroscopy (LIBS) data to enable rapid and accurate classification. However, for complex matrices like uranium (U) slag, individual feature selection methods often struggle to capture the full spectrum of discriminative information, leading to suboptimal model robustness. To address this, a weighted feature fusion (WFF) strategy is proposed for the first time to achieve high-precision classification of U slag. This strategy generates a comprehensive importance metric by performing a weighted linear fusion of normalized scores derived from random forest (RF), least absolute shrinkage and selection operator (LASSO), and mutual information (MI). The influence of diverse weight configurations on support vector machine (SVM), linear discriminant analysis (LDA), and k-nearest neighbors (KNN) models was systematically investigated using LIBS spectra from 23 U slag samples. The results demonstrate that the WFF strategy effectively reconciles the complementary strengths of the baseline methods—leveraging the dominant discriminative power of RF, the sparse linear features of LASSO, and the nonlinear associations from MI. Under the optimal weight configuration (RF: LASSO: MI = 0.5: 0.2: 0.3), the LDA model achieved a peak F1-score of 97.09%, significantly outperforming the best single-method approach (RF-LDA, 94.15%). The proposed strategy exhibits superior generalization and successfully mitigates the adaptation limitations typically observed when specific models are paired with individual selection methods. This study provides a novel, flexible, and interpretable feature engineering solution, offering critical methodological support for the field-deployable monitoring and resource utilization of nuclear-related solid waste.